RESEARCH FRONTIER
Daily selections of AI and AI agent research, including preprints, conference papers and research reports. Entries are grouped by edition, with original publication dates and limitations preserved.
Updated 2026-10-17
Rewrites dialogue into atomic memories and combines codebook associations with vector retrieval. Authors report accuracy and construction-efficiency gains on two memory benchmarks; scope remains the tested settings. Authors state acceptance at NeurIPS 2026.
Uses teacher–student progress gaps to adaptively penalize stalled trajectories. Authors report success-rate and task-efficiency improvements across eight small models; results do not establish universal task coverage. Authors state acceptance at NeurIPS 2026.
Group holdouts and paired interventions separate predictable labels from guard-decision changes. Some formatting cues affect false positives, with inconsistent effects across models. Authors state acceptance at the NeurIPS 2026 TAE and AIWILD workshops.
Combines hierarchical supervised tuning with windowed interaction reinforcement learning to credit multi-step experimental decisions. Authors report improved medal rates on MLE-Bench Lite and test task transfer. These are preprint benchmark results.
A builder reorganizes documents into hierarchical files and references, trained through downstream answer accuracy. Authors report gains on two benchmarks; upfront construction costs require repeated queries to amortize. Findings remain preprint results.
Trains agents with verifiable desktop tasks and action-step-balanced reinforcement learning. Authors report improved OSWorld-Verified success and test Windows transfer. Preprint benchmark results do not establish reliability across arbitrary desktop tasks.
Anthropic publishes code linking five NIST prime-curve seeds to lrand48 and historical timestamps. This does not break ECDSA; some base points and candidate-selection rules remain unexplained.
The study loops natural-language specifications through code generation, correctness checks and resource scoring across eight quantum-circuit tasks. The authors report new qubit–gate trade-offs and some formal verification. These are algorithmic resource estimates, not demonstrations on practical quantum hardware.
Surveys compare roughly 11,000 US adults with Anthropic staff, finding different expectations for AI capabilities and job automation. The study records beliefs and associations; it does not validate those forecasts or establish causal effects on the economy.
Internal employment, tool-use and output data plus seventeen interviews describe changing hiring and agent supervision. Tool adoption was not randomized, code volume is not quality, and one rapidly growing AI company cannot represent the wider labor market.
VERID tracks robot and laboratory state, checks physical constraints before execution, and supports iterative design with experimental feedback. The prototypes retain human authorization and supervision. Case studies do not establish fully autonomous laboratories or clinical effectiveness.
This preprint describes autonomous post-training on real GPUs, with frozen evaluation, job management and research records. Results are specific to its models and experimental setup, rather than evidence of general autonomous science.
An audit of ten runtimes motivates a gateway that renders canonical actions and executes only the stored, signed call. Experiments reduce substitution risks, but approval flooding remains effective and confirmations cannot block actions already permitted by policy.
The paper studies how HNSW graph structure can amplify document poisoning in a two-million-document Wikipedia index. It evaluates retrieval and downstream generation risks. Results are bounded by the tested index, search budgets and defense configurations.
Researchers translate user instructions while holding 1,220 tool-call steps, observations and labels fixed, then test nineteen guards. Several guards change verdicts beyond baseline noise in less common languages. This measures language sensitivity, rather than real-world attack success. Authors list acceptance to the NeurIPS 2026 AIWILD workshop, not the main conference.
MCP-FSD screens 21,097 servers across nine registries, combining contract analysis with paired dynamic execution. It confirms 150 vulnerable instances under tested sandbox policies. Screening proportions do not measure actual compromise across every MCP deployment.
The study combines independently developed specialist agents and models incentives for specialization. Their combined system reaches 70% on GAIA. These preliminary results concern the paper’s benchmark and development budgets, rather than all deployed agent teams.
Cross-Interface Policy Regularization aligns action distributions across two equivalent interfaces during post-training. Experiments report improved robustness to renamed actions and changed notation in supervised and reinforcement learning settings; the evidence covers the tested interface perturbations.
The authors periodically reset agent sessions and candidate histories while retaining the best solution. Four tasks in heuristic programming, kernel optimization and model training show improved search. The findings do not establish benefits across all scientific discovery settings.
Assembly and critic agents coordinate geometric planning with contact-dynamics feedback, while a skill library reuses simulator experience. Evaluations cover rigid, articulated and deformable objects in Isaac Sim; the reported improvements are simulation results, not physical robot validation.
A staged agent workflow creates query accelerators over existing database tables, maintaining auxiliary state during writes. It distinguishes differential testing from machine-checked proofs. Reported speedups apply to specified TPC-H workloads and update rates, not arbitrary database queries.
This preprint uses scientists’ unpublished research directions for evaluation. None of 135 follow-up papers generated by nine agents meets every condition, highlighting limits in research integrity and verification.
FEP expresses in-context nonparametric event prediction through attention, with robust embeddings and time normalization for heterogeneous logs. Pretraining on 696 datasets is evaluated without parameter adaptation on 50 unseen datasets; competitiveness varies across systems and prediction tasks.
YUBI-Real2Sim combines dual-fisheye capture, Gaussian scene reconstruction and physical simulation, reproducing real-robot policy ordering in eight of nine scene-task comparisons. Artifact interventions expose task sensitivity, while image-quality scores do not yet yield a generally reliable rule for accepting reconstructed scenes.
Box-WAM replaces dense future video with oriented bounding-box trajectories for task-relevant objects, jointly predicting actions and motion. Authors report reduced computation and improved action prediction in robotic manipulation; broader validation is needed where compact boxes may omit consequential scene details.
TacitVLA retains reasoning supervision while its action expert ignores reasoning generated for the current call; episodic entries can be produced during action execution. Four manipulation and driving benchmarks compare latency and performance. Controlled benchmark results do not establish real-world driving safety.
The study analyzes stale memories that replace multi-step reasoning, proposing likelihood-ratio monitoring and certificates restarted after change. Experiments span five models, edited-fact questions and verified code tasks. Its safety bounds depend on available verified outcomes and the specified information setting.
The paper derives mean-square benefits of layer-wise compensation under specific linear-network and rounding-noise assumptions, then constructs failures outside them. Experiments on two 8B models show that lowering calibration loss does not reliably improve unseen text, cautioning against universal guarantees.
ECHO separates introduced errors, inherited consequences and corrections using a frozen judge, counts shared consequences once and limits rewards for repairing self-induced mistakes. Six tool-use benchmarks show improvements at 4B and 8B scales; responsibility labels still depend on the judge.
Across 485 scenarios, authorized information can propagate to unauthorized agents through routine coordination. RECONS sanitizes messages according to recipient and execution context, reducing disclosure while preserving most authorized sharing. Residual leaks remain substantial, with detection coverage the dominant bottleneck.
An oracle-integrity audit combines state invariants and known-answer controls to identify a scoring defect that mutates benchmark state. Replay correction lowers measured success from 94.3% to 65.5% across 4,320 episodes. Portability fails when another benchmark cannot reconstruct reference terminal states.
Dream2Discern perturbs representation operators to create difficult latent variants of recorded trajectory prefixes, training a runtime outcome auditor without collecting new executions. Authors report over 90% weighted trajectory accuracy on AFTraj-2K and StepShield; evidence remains limited to these recorded benchmarks.
MoMaTune first models unlabeled serialized records, optionally matching partial views of each record, then fine-tunes target prediction. Eight public datasets show gains in numerical sensitivity and downstream representations. The evidence concerns record-based tasks, not general mathematical reasoning capability.
Sequent-Prover agents write SMT-LIB programs, consult Z3 and repair formalizations. Filtered trajectories support distillation and preference training, improving accuracy and reducing solver calls across 37 benchmarks. The evaluation distinguishes executable success from faithfully representing the original problem.
RECAP calibrates quantized reasoning models using long historical prefixes and rotating short reconstruction blocks, weighting cache errors by their attention effects. Authors report gains across several low-bit settings; benchmark improvements do not guarantee faithful reasoning for arbitrary long contexts.
LeHami replaces an unconstrained JEPA predictor with action-conditioned Hamiltonian dynamics integrated by a symplectic step, preserving latent phase-space volume. Visual pendulum experiments improve long-horizon velocity prediction; validation remains concentrated on conservative and actuated pendulum environments rather than complex real-world systems.
Co-DSL turns user descriptions and feedback into reusable spatial operations. A language model composes a concept graph while a solver finds satisfying layouts. Experiments span five layout domains; effectiveness depends on available concepts and constraints rather than unrestricted design understanding.
A customer-support case study combines a portable instruction skill, lightweight tools and a document navigation layer. Authors report strong workflow agreement in a curated corpus with human review, but roughly 90-second cases are slower than their production assistant; broader deployment remains untested. Authors list acceptance to the EMNLP 2026 Industry Track.
SceneForesee pairs observed and future scenes with shared entity identities. EvoSee jointly predicts trajectories and events to support temporal reasoning, improving the authors’ SceneForesee and Dyn-Bench evaluations. These benchmark gains do not establish reliable prediction of arbitrary future scenes.
TETPO tempers importance weights until each prompt group meets a target effective sample size, limiting domination by stale-policy trajectories. Authors derive dispersion bounds and evaluate mathematical reasoning under substantial policy lag; the analysis remains conditional on its bridge-distribution assumptions.
KV4-RL combines INT4 cache quantization with INT8 protection for attention sinks and recent context. Qwen3 experiments report faster asynchronous reinforcement-learning iterations with accuracy close to BF16 rollout training under a shared evaluator; results are limited to the tested configurations.
Executable task contracts specify defining facts, permitted edits and completion conditions before an agent runs. Rescoring 14,336 stored runs reveals that embedding similarity can admit task-changing edits. The contribution improves evaluation validity rather than establishing general agent safety.
MisVisAgentBench provides 280 executable task instances paired by clean or misleading chart evidence. Across thirteen agents, average success falls by over 34 percentage points; this measures a specific benchmark, not every real website or visualization. Authors list acceptance to EMNLP 2026.
Reference-answer likelihood guides branching at intermediate reasoning and tool-use steps, with corrected tree backups assigning turn-level advantages. Gains span seven question-answering benchmarks and three Qwen backbones; reliability on arbitrary open-ended tasks is not demonstrated.
A stratified analysis of 26,999 pull requests examines combined application and infrastructure changes, finding delayed reviews and unreviewed merges. The observational causal interpretation depends on conditional-exchangeability assumptions and does not eliminate all potential confounding.
The framework drafts, predicts feedback and revises internally, combining next-user-response supervision with GRPO. Reported gains cover collaborative mathematics, coding and writing tasks; these task-specific results do not establish human-like mental-state understanding.
Temporal discovery recovers nearby overlooked messages, while evidence forests track state transitions and compress low-information spans for answering. Improvements are reported on two long-term multi-party benchmarks; stability in ongoing real-world conversations remains to be validated.
Qwen3-1.7B predicts continuous visual tokens before matched image–text instruction tuning. Reported image and video scores improve while average text performance is retained; evidence is specific to this model and training setup, not a general scaling law.
DAG-EDA combines an intent graph with a knowledge graph linking domain concepts to dataset variables, supporting branching and backtracking. The paper presents a usage scenario and study design, not completed user-study evidence of improved analytical performance.
A four-pass compiler creates deterministic gates from written policies and checks tool-schema compatibility. Customer-service and injection benchmarks show fewer violations; protection depends on policy coverage and bindings, leaving actions outside governed tool calls beyond its scope.
The benchmark pairs 1,373 instructions and trajectories to expose hidden constraint violations across applications. Eleven judges are evaluated; the strongest agentic configuration reaches 80.9% balanced accuracy on the AH subset, leaving substantial room for misjudgment.
Eighty-two unresolved mathematics and theoretical-physics problems include context and prior progress. Four models independently judge submissions without reference solutions; reported solve rates are model assessments and should not be treated as independently certified mathematical discoveries.
Repeatedly selecting instruction changes on small reused evaluation sets exaggerates improvements in Qwen experiments. Independent held-out testing exposes the gap; the tested acceptance rules do not consistently resolve it, and estimates still carry substantial uncertainty.
Score-rank probabilities enable sample-only policy evaluation and uncertainty-aware sampling-budget selection. Synthetic tests and GSM8K support the estimator; statistical guarantees depend on the paper’s assumptions and do not guarantee error-free deployed responses.
A latent world model with a diffusion decoder supports anomaly signals, clustering and action–state consistency inspection. This qualitative case study uses CS:GO gameplay; quantitative curation metrics and downstream training gains are explicitly left untested.
A continuously evolving urban simulator evaluates both completion and safety events across eight vision-language models. Results show task success can conceal latency-related risk; simulated collision measurements should not be read as predictions of real-world accident rates.
A recurrent world model joins bird’s-eye exploration memory with task-aware spatial values for imagined rollouts. Reported success is 23.8% on UAV-ON and 19.9% in unseen scenes; these benchmark results do not establish reliable real-city autonomy.
Skill-specific adapters and a question-conditioned router reduce interference in audio understanding. The authors report average accuracy gains of 2.5–4.1 points across three models and three benchmarks; broader transfer remains untested.
Heterogeneous social-network simulations separate assigned personas from underlying models. Engagement depends more on the base model in these experiments, with lexical analyses exploring possible explanations; simulated interactions do not establish effects in human social networks. Authors note a spotlight presentation at a COLM 2026 social-simulation workshop.
An action-reconstruction objective helps world-model representations retain unstable but controllable directions. The paper proves results for linear systems and tests nonlinear visual-control tasks, rather than establishing stability guarantees for arbitrary robots.
Dynamic example selection and step-level gating aim to reduce gradient conflict and overwriting in LoRA updates. Results cover three backbones and seven instruction datasets; effectiveness across arbitrary architectures or distributions is not established.
Across 300 tasks, agents implement and refine checkers using repository and compiler feedback. Twenty-one model–harness configurations average 32.30% Pass@1, exposing difficulty on this benchmark without measuring complete security-audit capability.
No research matches this tag and date range. Choose another tag or all dates.